Information Extraction – why Google doesn’t even come close

1. Diana Maynard Natural Language Processing Group University of Sheffield, UK BCS meeting, 25 September 2003 Information Extraction – why Google doesn’t even come close 1()

2. Information Extraction and Information Retrieval The MUSE system for Named Entity Recognition Multilingual MUSE Future directions Outline 2()

3. IE pulls facts and structured information from the content of large text collections (usually corpora). You analyse the facts. IR pulls documents from large text collections (usually the Web) in response to specific keywords or queries. You analyse the documents. IE is not IR 3()

4. With traditional query engines, getting the facts can be hard and slow Where has the Queen visited in the last year? Which places on the East Coast of the US have had cases of West Nile Virus? Which search terms would you use to get this kind of information? IE would return information in a structured way IR would return documents containing the relevant information somewhere (if you were lucky) IE for Document Access 4()

5. IE returns knowledge at a much deeper level than IR Constructing a database through IE and linking it back to the documents can provide a valuable alternative search tool. Even if results are not always accurate, they can be valuable if linked back to the original text IE as an alternative to IR 5()

6. For access to news identify major relations and event types (e.g. within foreign affairs or business news) For access to scientific reports identify principal relations of a scientific subfield (e.g. pharmacology, genomics) When would you use IE? 6()

7. Aims to find out how companies report about health and safety information Answers questions such as: “how many members of staff died or had accidents in the last year?” “is there anyone responsible for health and safety” “what measures have been put in place to improve health and safety in the workplace?” Application 1 – HaSIE 7()

8. Identification of such information is too time-consuming and arduous to be done manually IR systems can’t cope with this because they return whole documents, which could be hundreds of pages System identifies relevant sections of each document, pulls out sentences about health and safety issues, and populates a database with relevant information HASIE 8()

9. Application 2: KIM Ontotext’s KIM query and results 9()

10. Application 3: Threat tracker 10()

11. Identification of proper names in texts, and their classification into a set of predefined categories of interest Persons Organisations (companies, government organisations, committees, etc) Locations (cities, countries, rivers, etc) Date and time expressions Various other types as appropriate What is Named Entity Recognition? 11()

12. NE provides a foundation from which to build more complex IE systems Relations between NEs can provide tracking, ontological information and scenario building Tracking (co-reference) “Dr Head, John, he” Ontologies “Manchester, CT” Scenario “Dr Head became the new director of Shiny Rockets Corp” Why is NE important 12()

13. Knowledge Engineering rule based developed by experienced language engineers make use of human intuition require only small amount of training data development can be very time consuming some changes may be hard to accommodate Learning Systems use statistics or other machine learning developers do not need LE expertise require large amounts of annotated training data some changes may require re-annotation of the entire training corpus Two kinds of approaches 13()

14. Variation of NEs – e.g. John Smith, Mr Smith, John. Ambiguity of NE types: John Smith (company vs. person) June (person vs. month) Washington (person vs. location) 1945 (date vs. time) Ambiguity between common words and proper nouns, e.g. “may” Basic Problems in NE 14()

15. Issues of style, structure, domain, genre etc. Punctuation, spelling, spacing, formatting Dept. of Computing and Maths Manchester Metropolitan University Manchester United Kingdom > Tell me more about Leonardo > Da Vinci More complex problems in NE 15()

16. System that recognises only entities stored in its lists (gazetteers). Advantages - Simple, fast, language independent, easy to retarget (just create lists) Disadvantages - collection and maintenance of lists, cannot deal with name variants, cannot resolve ambiguity List lookup approach - baseline 16()

17. Internal evidence – names often have internal structure. These components can be either stored or guessed, e.g. location: Cap. Word + {City, Forest, Center, River} e.g. Sherwood Forest Cap. Word + {Street, Boulevard, Avenue, Crescent, Road} e.g. Portobello Street Shallow Parsing Approach (internal structure) 17()

18. Ambiguously capitalised words (first word in sentence)[All American Bank] vs. All [State Police] Semantic ambiguity "John F. Kennedy" = airport (location) "Philip Morris" = organisation Structural ambiguity [Cable and Wireless] vs. [Microsoft] and [Dell] [Center for Computational Linguistics] vs. message from [City Hospital] for [John Smith] Problems with the shallow parsing approach 18()

19. Use of context-based patterns is helpful in ambiguous cases "David Walton" and "Goldman Sachs" are indistinguishable But with the phrase "David Walton of Goldman Sachs" and the Person entity "David Walton" recognised, we can use the pattern "[Person] of [Organization]" to identify "Goldman Sachs“ correctly. Shallow Parsing Approach with Context 19()

20. Use KWIC index and concordancer to find windows of context around entities Search for repeated contextual patterns of either strings, other entities, or both Manually post-edit list of patterns, and incorporate useful patterns into new rules Repeat with new entities Identification of Contextual Information (1) 20()

21. [PERSON] earns [MONEY] [PERSON] joined [ORGANIZATION] [PERSON] left [ORGANIZATION] [PERSON] joined [ORGANIZATION] as [JOBTITLE] [ORGANIZATION]'s [JOBTITLE] [PERSON] [ORGANIZATION] [JOBTITLE] [PERSON] the [ORGANIZATION] [JOBTITLE] part of the [ORGANIZATION] [ORGANIZATION] headquarters in [LOCATION] price of [ORGANIZATION] sale of [ORGANIZATION] investors in [ORGANIZATION] [ORGANIZATION] is worth [MONEY] [JOBTITLE] [PERSON] [PERSON], [JOBTITLE] Examples of semantic patterns 21()

22. Automatic collection of context words with particular features Collect e.g. all verbs preceding a Person annotation (from training data) Sort verb list by frequency and use cut off threshold (optional) Verbs can then be used to search for new Persons Repeat procedure with newly identified Persons Contextual Patterns (2) 22()

23. An IE system developed within GATE Performs NE and coreference on different text types and genres Uses knowledge engineering approach with hand-crafted rules Performance rivals that of machine learning methods Easily adaptable MUSE – MUlti-Source Entity Recognition 23()

24. Document format and genre analysis Tokenisation Sentence splitting POS tagging Gazetteer lookup Semantic grammar Orthographic coreference Nominal and pronominal coreference MUSE Modules 24()

25. Rather than have a fixed chain of processing resources, choices can be made automatically about which modules to use Texts are analysed for certain identifying features which are used to trigger different modules For example, texts with no case information may need different POS tagger or gazetteer lists Not all modules are language-dependent, so some can be reused directly Switching Controller 25()

26. MUSE has been adapted to deal with different languages Currently systems for English, French, German, Romanian, Bulgarian, Russian, Cebuano, Hindi, Chinese, Arabic Separation of language-dependent and language-independent modules and sub-modules Annotation projection experiments Multilingual MUSE 26()

27. Adaptation to an unknown language in a very short timespan Cebuano: Latin script, capitalisation, words are spaced Few resources and little work already done Medium difficulty Hindi: Non-Latin script, different encodings used, no capitalisation, words are spaced Many resources available Medium difficulty IE in Surprise Languages 27()

28. Extensive support for non-Latin scripts and text encodings, including conversion utilities Automatic recognition of encoding Occupied up to 2/3 of the TIDES Hindi effort Bilingual dictionaries Annotated corpus for evaluation Internet resources for gazetteer list collection (e.g., phone books, yellow pages, bi-lingual pages) What does multilingual NE require? 28()

29. Editing Multilingual Data • GATE Unicode Kit (GUK) • Complements Java’s facilities • Support for defining Input Methods (IMs) • currently 30 IMs for 17 languages • Pluggable in other applications (e.g. JEdit) 29()

30. Processing Multilingual Data All processing, visualisation and editing tools use GUK 30()

31. ML methods and robust IE systems mean high quality results can be achieved fast Fast adaptation to new languages is the focus of much current work – especially languages such as Arabic, Chinese, Japanese… So what does the future hold for IE? State of the art in IE research 31()

32. Tools for semantic web Hierarchical NE recognition Need for IE in bioinformatics and medicine is becoming increasingly evident Cross fertilisation of IE and IR , eg. For Question Answering Collaboration between fields of IE and computational terminology The future of IE 32()